First, please follow frk_anonymous/README.md to install our C++ implementations of FRK.

1. To reproduce results of SVMs in Section 4, run frk_mp.py or *_mp.py to calcurate gram matrices for each kernel or distance. Then, run run_svm.py to get scores. For RNNs, run demo_lstm.py or demo_gru.py

Preparation:

This library is a C++11 implementation of Frechet Kernel (FRK). You can build a Python module using Pybind11.

The library builds a Python module from src/frk.cpp using Pybind11, so you have to install Pybind11. Following the official document, you can install it from the source files as follows.

git clone https://github.com/pybind/pybind11.git
cd pybind11
mkdir build
cd build
cmake ..
make check -j4
make install

Then, you can build pyfrk module through setup.py in the following command.

pip install .

If necessary, set the include path for Pybind11 on line 19 of setup.py.

pyfrk module has the following classes and functions:

• pyfrk.MatrixF64: A matrix of Float64.
  • __init__(nrows, ncols) reserves the matrix size.
  • set(i,j,e) sets value e in row 0<=i<nrows, column 0<=j<ncols of the matrix.
• pyfrk.Kernel: Class for computing Frechet Kernel values.
  • __init__(maxsize) initializes the class using the maximum length of input trajectories, taking O(maxsize^2) time.
  • compute(emat,nsamples,beta,diag_wgt,seed) computes the kernel value from a distance matrix emat=pyfrk.MatrixF64 (i.e., FRK1).
    • nsamples is the number of samples, beta is a parameter for smooth-min function, diag_wgt is a weight of the probability of transition to the diagonal, seed is a seed value for uniform distribution. Setting diag_wgt=1.0 indicates random sampling on uniform distribution. Setting diag_wgt>1.0 increases the probability of transition to the diagonal.

sample/sample.py

#!/usr/bin/env python3

import pyfrk
import random
import math

random.seed(114514)

N = 10  # Number of sequence
L = 20  # Length of each sequence

# Generate random N sequences
seqs = [[random.random() for i in range(L)] for j in range(N)]

# Parameters for Sampling Alignment Kernel
SAMPLES = 10
BETA = 1.0
DIAG_WGT = 1.0 # Diagonal weight for random walk
SEED = random.randint(0, (1 << 64) - 1) # Random seed for random walk
GAMMA = 1.0 # for e_ij

# Engine of Sampling Alignment Kernel
ker = pyfrk.Kernel(L) # L is the maximum length of sequence considered

for x in range(N):
    for y in range(x, N):
        seq_x = seqs[x]
        seq_y = seqs[y]

        # Make e_ij matrix, i.e., \phi(xi,yj) = exp(-d(xi,yi)/\gamma)
        emat = pyfrk.MatrixF64(len(seq_x), len(seq_y))
        for i, xv in enumerate(seq_x):
            for j, yv in enumerate(seq_y):
                emat.set(i, j, math.exp(-abs(xv - yv)/GAMMA))

        try:
            ans = ker.compute(emat=emat, nsamples=SAMPLES, beta=BETA, diag_wgt=DIAG_WGT, seed=SEED)
            print(f'{x}, {y} => {ans:g}')
        except RuntimeError as e:
            print('Exception:', e)